drivers/net/ethernet/mellanox/mlx4/icm.c

   1 /*
   2  * Copyright (c) 2005, 2006, 2007, 2008 Mellanox Technologies. All rights reserved.
   3  * Copyright (c) 2006, 2007 Cisco Systems, Inc.  All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/errno.h>
  35 #include <linux/mm.h>
  36 #include <linux/scatterlist.h>
  37 #include <linux/slab.h>
  38
  39 #include <linux/mlx4/cmd.h>
  40
  41 #include "mlx4.h"
  42 #include "icm.h"
  43 #include "fw.h"
  44
  45 /*
  46  * We allocate in as big chunks as we can, up to a maximum of 256 KB
  47  * per chunk.
  48  */
  49 enum {
  50         MLX4_ICM_ALLOC_SIZE     = 1 << 18,
  51         MLX4_TABLE_CHUNK_SIZE   = 1 << 18
  52 };
  53
  54 static void mlx4_free_icm_pages(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
  55 {
  56         int i;
  57
  58         if (chunk->nsg > 0)
  59                 pci_unmap_sg(dev->pdev, chunk->mem, chunk->npages,
  60                              PCI_DMA_BIDIRECTIONAL);
  61
  62         for (i = 0; i < chunk->npages; ++i)
  63                 __free_pages(sg_page(&chunk->mem[i]),
  64                              get_order(chunk->mem[i].length));
  65 }
  66
  67 static void mlx4_free_icm_coherent(struct mlx4_dev *dev, struct mlx4_icm_chunk *chunk)
  68 {
  69         int i;
  70
  71         for (i = 0; i < chunk->npages; ++i)
  72                 dma_free_coherent(&dev->pdev->dev, chunk->mem[i].length,
  73                                   lowmem_page_address(sg_page(&chunk->mem[i])),
  74                                   sg_dma_address(&chunk->mem[i]));
  75 }
  76
  77 void mlx4_free_icm(struct mlx4_dev *dev, struct mlx4_icm *icm, int coherent)
  78 {
  79         struct mlx4_icm_chunk *chunk, *tmp;
  80
  81         if (!icm)
  82                 return;
  83
  84         list_for_each_entry_safe(chunk, tmp, &icm->chunk_list, list) {
  85                 if (coherent)
  86                         mlx4_free_icm_coherent(dev, chunk);
  87                 else
  88                         mlx4_free_icm_pages(dev, chunk);
  89
  90                 kfree(chunk);
  91         }
  92
  93         kfree(icm);
  94 }
  95
  96 static int mlx4_alloc_icm_pages(struct scatterlist *mem, int order,
  97                                 gfp_t gfp_mask, int node)
  98 {
  99         struct page *page;
 100
 101         page = alloc_pages_node(node, gfp_mask, order);
 102         if (!page) {
 103                 page = alloc_pages(gfp_mask, order);
 104                 if (!page)
 105                         return -ENOMEM;
 106         }
 107
 108         sg_set_page(mem, page, PAGE_SIZE << order, 0);
 109         return 0;
 110 }
 111
 112 static int mlx4_alloc_icm_coherent(struct device *dev, struct scatterlist *mem,
 113                                     int order, gfp_t gfp_mask)
 114 {
 115         void *buf = dma_alloc_coherent(dev, PAGE_SIZE << order,
 116                                        &sg_dma_address(mem), gfp_mask);
 117         if (!buf)
 118                 return -ENOMEM;
 119
 120         sg_set_buf(mem, buf, PAGE_SIZE << order);
 121         BUG_ON(mem->offset);
 122         sg_dma_len(mem) = PAGE_SIZE << order;
 123         return 0;
 124 }
 125
 126 struct mlx4_icm *mlx4_alloc_icm(struct mlx4_dev *dev, int npages,
 127                                 gfp_t gfp_mask, int coherent)
 128 {
 129         struct mlx4_icm *icm;
 130         struct mlx4_icm_chunk *chunk = NULL;
 131         int cur_order;
 132         int ret;
 133
 134         /* We use sg_set_buf for coherent allocs, which assumes low memory */
 135         BUG_ON(coherent && (gfp_mask & __GFP_HIGHMEM));
 136
 137         icm = kmalloc_node(sizeof(*icm),
 138                            gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN),
 139                            dev->numa_node);
 140         if (!icm) {
 141                 icm = kmalloc(sizeof(*icm),
 142                               gfp_mask & ~(__GFP_HIGHMEM | __GFP_NOWARN));
 143                 if (!icm)
 144                         return NULL;
 145         }
 146
 147         icm->refcount = 0;
 148         INIT_LIST_HEAD(&icm->chunk_list);
 149
 150         cur_order = get_order(MLX4_ICM_ALLOC_SIZE);
 151
 152         while (npages > 0) {
 153                 if (!chunk) {
 154                         chunk = kmalloc_node(sizeof(*chunk),
 155                                              gfp_mask & ~(__GFP_HIGHMEM |
 156                                                           __GFP_NOWARN),
 157                                              dev->numa_node);
 158                         if (!chunk) {
 159                                 chunk = kmalloc(sizeof(*chunk),
 160                                                 gfp_mask & ~(__GFP_HIGHMEM |
 161                                                              __GFP_NOWARN));
 162                                 if (!chunk)
 163                                         goto fail;
 164                         }
 165
 166                         sg_init_table(chunk->mem, MLX4_ICM_CHUNK_LEN);
 167                         chunk->npages = 0;
 168                         chunk->nsg    = 0;
 169                         list_add_tail(&chunk->list, &icm->chunk_list);
 170                 }
 171
 172                 while (1 << cur_order > npages)
 173                         --cur_order;
 174
 175                 if (coherent)
 176                         ret = mlx4_alloc_icm_coherent(&dev->pdev->dev,
 177                                                       &chunk->mem[chunk->npages],
 178                                                       cur_order, gfp_mask);
 179                 else
 180                         ret = mlx4_alloc_icm_pages(&chunk->mem[chunk->npages],
 181                                                    cur_order, gfp_mask,
 182                                                    dev->numa_node);
 183
 184                 if (ret) {
 185                         if (--cur_order < 0)
 186                                 goto fail;
 187                         else
 188                                 continue;
 189                 }
 190
 191                 ++chunk->npages;
 192
 193                 if (coherent)
 194                         ++chunk->nsg;
 195                 else if (chunk->npages == MLX4_ICM_CHUNK_LEN) {
 196                         chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 197                                                 chunk->npages,
 198                                                 PCI_DMA_BIDIRECTIONAL);
 199
 200                         if (chunk->nsg <= 0)
 201                                 goto fail;
 202                 }
 203
 204                 if (chunk->npages == MLX4_ICM_CHUNK_LEN)
 205                         chunk = NULL;
 206
 207                 npages -= 1 << cur_order;
 208         }
 209
 210         if (!coherent && chunk) {
 211                 chunk->nsg = pci_map_sg(dev->pdev, chunk->mem,
 212                                         chunk->npages,
 213                                         PCI_DMA_BIDIRECTIONAL);
 214
 215                 if (chunk->nsg <= 0)
 216                         goto fail;
 217         }
 218
 219         return icm;
 220
 221 fail:
 222         mlx4_free_icm(dev, icm, coherent);
 223         return NULL;
 224 }
 225
 226 static int mlx4_MAP_ICM(struct mlx4_dev *dev, struct mlx4_icm *icm, u64 virt)
 227 {
 228         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM, icm, virt);
 229 }
 230
 231 static int mlx4_UNMAP_ICM(struct mlx4_dev *dev, u64 virt, u32 page_count)
 232 {
 233         return mlx4_cmd(dev, virt, page_count, 0, MLX4_CMD_UNMAP_ICM,
 234                         MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
 235 }
 236
 237 int mlx4_MAP_ICM_AUX(struct mlx4_dev *dev, struct mlx4_icm *icm)
 238 {
 239         return mlx4_map_cmd(dev, MLX4_CMD_MAP_ICM_AUX, icm, -1);
 240 }
 241
 242 int mlx4_UNMAP_ICM_AUX(struct mlx4_dev *dev)
 243 {
 244         return mlx4_cmd(dev, 0, 0, 0, MLX4_CMD_UNMAP_ICM_AUX,
 245                         MLX4_CMD_TIME_CLASS_B, MLX4_CMD_NATIVE);
 246 }
 247
 248 int mlx4_table_get(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj)
 249 {
 250         u32 i = (obj & (table->num_obj - 1)) /
 251                         (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
 252         int ret = 0;
 253
 254         mutex_lock(&table->mutex);
 255
 256         if (table->icm[i]) {
 257                 ++table->icm[i]->refcount;
 258                 goto out;
 259         }
 260
 261         table->icm[i] = mlx4_alloc_icm(dev, MLX4_TABLE_CHUNK_SIZE >> PAGE_SHIFT,
 262                                        (table->lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 263                                        __GFP_NOWARN, table->coherent);
 264         if (!table->icm[i]) {
 265                 ret = -ENOMEM;
 266                 goto out;
 267         }
 268
 269         if (mlx4_MAP_ICM(dev, table->icm[i], table->virt +
 270                          (u64) i * MLX4_TABLE_CHUNK_SIZE)) {
 271                 mlx4_free_icm(dev, table->icm[i], table->coherent);
 272                 table->icm[i] = NULL;
 273                 ret = -ENOMEM;
 274                 goto out;
 275         }
 276
 277         ++table->icm[i]->refcount;
 278
 279 out:
 280         mutex_unlock(&table->mutex);
 281         return ret;
 282 }
 283
 284 void mlx4_table_put(struct mlx4_dev *dev, struct mlx4_icm_table *table, u32 obj)
 285 {
 286         u32 i;
 287         u64 offset;
 288
 289         i = (obj & (table->num_obj - 1)) / (MLX4_TABLE_CHUNK_SIZE / table->obj_size);
 290
 291         mutex_lock(&table->mutex);
 292
 293         if (--table->icm[i]->refcount == 0) {
 294                 offset = (u64) i * MLX4_TABLE_CHUNK_SIZE;
 295                 mlx4_UNMAP_ICM(dev, table->virt + offset,
 296                                MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 297                 mlx4_free_icm(dev, table->icm[i], table->coherent);
 298                 table->icm[i] = NULL;
 299         }
 300
 301         mutex_unlock(&table->mutex);
 302 }
 303
 304 void *mlx4_table_find(struct mlx4_icm_table *table, u32 obj,
 305                         dma_addr_t *dma_handle)
 306 {
 307         int offset, dma_offset, i;
 308         u64 idx;
 309         struct mlx4_icm_chunk *chunk;
 310         struct mlx4_icm *icm;
 311         struct page *page = NULL;
 312
 313         if (!table->lowmem)
 314                 return NULL;
 315
 316         mutex_lock(&table->mutex);
 317
 318         idx = (u64) (obj & (table->num_obj - 1)) * table->obj_size;
 319         icm = table->icm[idx / MLX4_TABLE_CHUNK_SIZE];
 320         dma_offset = offset = idx % MLX4_TABLE_CHUNK_SIZE;
 321
 322         if (!icm)
 323                 goto out;
 324
 325         list_for_each_entry(chunk, &icm->chunk_list, list) {
 326                 for (i = 0; i < chunk->npages; ++i) {
 327                         if (dma_handle && dma_offset >= 0) {
 328                                 if (sg_dma_len(&chunk->mem[i]) > dma_offset)
 329                                         *dma_handle = sg_dma_address(&chunk->mem[i]) +
 330                                                 dma_offset;
 331                                 dma_offset -= sg_dma_len(&chunk->mem[i]);
 332                         }
 333                         /*
 334                          * DMA mapping can merge pages but not split them,
 335                          * so if we found the page, dma_handle has already
 336                          * been assigned to.
 337                          */
 338                         if (chunk->mem[i].length > offset) {
 339                                 page = sg_page(&chunk->mem[i]);
 340                                 goto out;
 341                         }
 342                         offset -= chunk->mem[i].length;
 343                 }
 344         }
 345
 346 out:
 347         mutex_unlock(&table->mutex);
 348         return page ? lowmem_page_address(page) + offset : NULL;
 349 }
 350
 351 int mlx4_table_get_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 352                          u32 start, u32 end)
 353 {
 354         int inc = MLX4_TABLE_CHUNK_SIZE / table->obj_size;
 355         int err;
 356         u32 i;
 357
 358         for (i = start; i <= end; i += inc) {
 359                 err = mlx4_table_get(dev, table, i);
 360                 if (err)
 361                         goto fail;
 362         }
 363
 364         return 0;
 365
 366 fail:
 367         while (i > start) {
 368                 i -= inc;
 369                 mlx4_table_put(dev, table, i);
 370         }
 371
 372         return err;
 373 }
 374
 375 void mlx4_table_put_range(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 376                           u32 start, u32 end)
 377 {
 378         u32 i;
 379
 380         for (i = start; i <= end; i += MLX4_TABLE_CHUNK_SIZE / table->obj_size)
 381                 mlx4_table_put(dev, table, i);
 382 }
 383
 384 int mlx4_init_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table,
 385                         u64 virt, int obj_size, u32 nobj, int reserved,
 386                         int use_lowmem, int use_coherent)
 387 {
 388         int obj_per_chunk;
 389         int num_icm;
 390         unsigned chunk_size;
 391         int i;
 392         u64 size;
 393
 394         obj_per_chunk = MLX4_TABLE_CHUNK_SIZE / obj_size;
 395         num_icm = (nobj + obj_per_chunk - 1) / obj_per_chunk;
 396
 397         table->icm      = kcalloc(num_icm, sizeof *table->icm, GFP_KERNEL);
 398         if (!table->icm)
 399                 return -ENOMEM;
 400         table->virt     = virt;
 401         table->num_icm  = num_icm;
 402         table->num_obj  = nobj;
 403         table->obj_size = obj_size;
 404         table->lowmem   = use_lowmem;
 405         table->coherent = use_coherent;
 406         mutex_init(&table->mutex);
 407
 408         size = (u64) nobj * obj_size;
 409         for (i = 0; i * MLX4_TABLE_CHUNK_SIZE < reserved * obj_size; ++i) {
 410                 chunk_size = MLX4_TABLE_CHUNK_SIZE;
 411                 if ((i + 1) * MLX4_TABLE_CHUNK_SIZE > size)
 412                         chunk_size = PAGE_ALIGN(size -
 413                                         i * MLX4_TABLE_CHUNK_SIZE);
 414
 415                 table->icm[i] = mlx4_alloc_icm(dev, chunk_size >> PAGE_SHIFT,
 416                                                (use_lowmem ? GFP_KERNEL : GFP_HIGHUSER) |
 417                                                __GFP_NOWARN, use_coherent);
 418                 if (!table->icm[i])
 419                         goto err;
 420                 if (mlx4_MAP_ICM(dev, table->icm[i], virt + i * MLX4_TABLE_CHUNK_SIZE)) {
 421                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 422                         table->icm[i] = NULL;
 423                         goto err;
 424                 }
 425
 426                 /*
 427                  * Add a reference to this ICM chunk so that it never
 428                  * gets freed (since it contains reserved firmware objects).
 429                  */
 430                 ++table->icm[i]->refcount;
 431         }
 432
 433         return 0;
 434
 435 err:
 436         for (i = 0; i < num_icm; ++i)
 437                 if (table->icm[i]) {
 438                         mlx4_UNMAP_ICM(dev, virt + i * MLX4_TABLE_CHUNK_SIZE,
 439                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 440                         mlx4_free_icm(dev, table->icm[i], use_coherent);
 441                 }
 442
 443         kfree(table->icm);
 444
 445         return -ENOMEM;
 446 }
 447
 448 void mlx4_cleanup_icm_table(struct mlx4_dev *dev, struct mlx4_icm_table *table)
 449 {
 450         int i;
 451
 452         for (i = 0; i < table->num_icm; ++i)
 453                 if (table->icm[i]) {
 454                         mlx4_UNMAP_ICM(dev, table->virt + i * MLX4_TABLE_CHUNK_SIZE,
 455                                        MLX4_TABLE_CHUNK_SIZE / MLX4_ICM_PAGE_SIZE);
 456                         mlx4_free_icm(dev, table->icm[i], table->coherent);
 457                 }
 458
 459         kfree(table->icm);
 460 }